Project summary

I am seeking USD 24,000 for four months of work on a practical economic question: how much human attention does it take to turn AI output into a result a research team can use? My ongoing Generative Economics programme studies how human capability and AI autonomy shape production. This grant supports one core paper and a new evaluation of checking costs, with a USD 12,000 minimum for a narrower three-month stage.

A team choosing an AI workflow needs to know more than how quickly it produces a draft. It needs the time required to find unsupported claims, correct them and decide that the result is fit for use. I will measure those costs on research tasks with traceable source records, then publish the evidence and a worksheet that lets a team compare model spending with the cost of its own review time.

The economic manuscripts and related data research already exist. The timed evaluation proposed here has not been conducted. Funding buys the next stage of work: model revision, new measurement, a reproducible public release and prospective publication expenses.

Why this project and why me

My research repeatedly encounters the gap between a usable record and an apparently convincing one. My recent sole-authored papers examine coordination in global AI governance in Policy Sciences and the interpretation of pharmaceutical authorisation records in Health Policy and Technology. Current coauthored work on corporate AI disclosure keeps source statements separate from judgments about whether a business change is actually attributable to AI. That work provides experience in building and checking evidence; it is not an already completed experiment on AI productivity.

I also teach generative AI for technology-management research and have delivered AI training to public-sector and healthcare audiences. These experiences help define relevant tasks and explain the findings. They do not imply that those organisations have agreed to participate or use the outputs.

Existing workplace and developer studies already examine AI productivity. My proposed contribution is narrower: a source-linked account of the human work between an AI answer and an accepted research result, connected to an economic model of scarce review capacity. The broader programme also studies learning and innovation participation, but this award has one bounded set of deliverables.

What I will do

I will revise the core model around candidate load, checking capacity and usable output. The empirical module will measure checking requirements within defined workflows. It will not, by itself, identify the full production function or the causal effect of human ability, organisational incentives or AI autonomy.

The full evaluation will use 120 tasks built from public research-funding records, beginning with NIH RePORTER. The three task types are extracting defined facts, summarising a project's stated aims and deciding whether a source supports a claim. Before generation, each task will have a source packet, reference answer, acceptance rule and an explicit unresolved category. Reference construction time will be recorded separately from review time.

Two AI systems will each use two workflows, with two runs per task: 960 outputs. The workflows receive the same source material; one returns a direct answer and the other must connect material claims to evidence and mark uncertainty. A balanced sample of at least 120 outputs will receive timed human review. Sampling rules, a review-time limit and error severity definitions will be fixed after the pilot and before the main evaluation. Model versions, prompts, usage and costs will be retained.

Review will record time to check and correct an answer, remaining material errors and whether it meets the acceptance rule within the time limit. A paid second reviewer will independently score a prespecified subset. Review order will be balanced; exposure to repeated versions of the same task will be logged and limited. I will report reviewer-specific timing and uncertainty at the task level, rather than count repeated outputs as independent tasks.

The main comparison is between the two AI-assisted workflows. There is no measured human-only control in this budget. The worksheet will report cost per accepted result and accepted results per recorded review hour for the timed sample, including the cost of unsuccessful reviews. Preparation, generation and review costs will remain separate. Any comparison with human-only work will be labelled as a sensitivity scenario, not an observed productivity gain.

What the grant will deliver

Month one covers the model revision, acceptance rules and a 30-task pilot. Months two and three cover the full evaluation and independent checks. Month four produces one working paper prepared for submission, the protocol and original code and annotations, and the cost-comparison worksheet with a short guide for research teams. The guide will show when an evidence requirement reduces correction work and when it merely makes an answer look more documented.

At USD 12,000 I will complete the 30-task pilot, 240 AI outputs and at least 40 timed reviews, with a focused revision of the same paper and a smaller public package. This is a viable pilot award, not funding for the full programme. At either level I will report after the pilot and at completion. Open code, task identifiers and permitted annotations will let others rerun the evaluation with later models; source text will be redistributed only where allowed.

How funding will be used

The USD 24,000 budget includes USD 12,000 for 160 hours of my research time at USD 75 per hour; USD 3,000 for 120 assistant hours at USD 25; USD 1,000 for 20 independent-review hours at USD 50; USD 1,000 for model access; USD 1,000 for data, storage and release; and up to USD 6,000 for submission fees and article processing charges for the project output. My commitment averages about ten hours per week over four months.

At the USD 12,000 minimum, the corresponding amounts are USD 6,000 for 80 hours of my time, USD 1,000 for 40 assistant hours, USD 500 for ten review hours, USD 500 for model access, USD 500 for data and release, and USD 3,500 for publication costs. Paid research time is retained at both levels. The larger award buys the full evaluation and additional model work; it does not simply increase the publication allowance.

Publication costs are prospective reserves, not journal quotations or a promise of acceptance. I will check institutional coverage and waivers, document actual charges, and seek approval to reallocate or return unused funds. The public working paper and permitted research materials will be released even if journal review continues beyond the project. Compensation figures are gross project costs.

Risks and boundaries

The tasks may be easier to check than real research, source records may be ambiguous, and reviewer experience may dominate differences between workflows. The pilot will expose these limits before expansion. I will report unresolved cases and findings that run against the model's motivating argument. A format that provides more citations may still contain wrong claims or take longer to review.

The intended users are small research teams and analysts choosing review procedures. I am not claiming confirmed partner demand, occupation-wide productivity effects or a measured long-run growth effect. The grant supports research and public research outputs, not commercial product development.

Researcher and other funding

I am Hana Kim, a contract research professor at KAIST's School of Business and Technology Management in South Korea. I hold a PhD in Technology Management, Economics and Policy and a master's degree in medicine from Seoul National University. I will lead the funded work from Korea. Assistants and independent reviewers will be paid from the grant and recruited once support is secured.

I have submitted an Emergent Ventures request for USD 30,000 for a six-month stage of the same programme. It overlaps with this proposal. If both attract support, I will disclose the overlap and agree a revised scope and budget before accepting awards; the same time, task construction or publication charge will not be funded twice. Additional proposals for founder histories or scientific reuse concern separate research questions and will have separate costs and a feasible combined schedule.

I am not currently receiving a separate research grant for the work proposed here. I participate in an institution-administered research institute project supported by the National Research Foundation of Korea (NRF). That participation is not presented here as a personal grant award. I will check the allocation of research time and expenses so that any costs covered by the institutional project are not also charged to this grant.

Recent publications

Kim, H. (2026). Coordinating fragmented global AI governance: Control, Rights, and Safety as a regime complex. Policy Sciences. Published online 23 September 2026. DOI: 10.1007/s11077-026-09632-w.

Kim, H. (2026). Source-record correlates of limited observed authorisation-holder diversity in marketed archive records with recorded exclusivity expiry. Health Policy and Technology, 101304. In press. DOI: 10.1016/j.hlpt.2026.101304.

Academic profile: Hana Kim at KAIST BTM

Publication links: Policy Sciences | Health Policy and Technology